Flow with LSD v1

blue-machines/Flow-with-LSD-v1 is a Gemma 3 270M two-head ONNX model (intent removed):

LID — sentence language (10 Indian languages)
LSD — language switch detection (10 languages + no_switch)

Derived from blue-machines/Floe-with-intent-classifer-v1 by stripping the intent head from the deploy ONNX graph. Encoder + LID remain INT8; LSD head stays FP32.

Quantization

Part	Precision
Gemma encoder	INT8
LID head	INT8
LSD head	FP32

Deploy size: ~258 MB (model.onnx).

Gate metrics (from parent Floe checkpoint)

Task	Metric	Value
LID	`lid_finetune_subset` acc	0.9944648823787505
LSD	`lsd_evaluation` acc	0.9704418789808917
LSD	`lsd_evaluation` no_switch	0.973753280839895

Files

File	Role
`model.onnx`	Deploy — LID + LSD only (no intent)
`model_fp32.onnx`	FP32 reference, LID + LSD only
`label_map.json`	LID / LSD label ids
`tokenizer.json`	Gemma tokenizer (max length 128)

Inference

import json
from pathlib import Path

import numpy as np
import onnxruntime as ort
from transformers import AutoTokenizer

MODEL_DIR = Path(".")  # or snapshot_download("blue-machines/Flow-with-LSD-v1")
MAX_LEN = 128

tokenizer = AutoTokenizer.from_pretrained(MODEL_DIR)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

with open(MODEL_DIR / "label_map.json", encoding="utf-8") as f:
    maps = json.load(f)

id2lid = {int(k): v for k, v in maps["head1_lid"]["id2label"].items()}
id2switch = {int(k): v for k, v in maps["head2_switch"]["id2label"].items()}

session = ort.InferenceSession(
    str(MODEL_DIR / "model.onnx"),
    providers=["CPUExecutionProvider"],
)

def predict(text: str) -> dict:
    enc = tokenizer(
        text,
        return_tensors="np",
        truncation=True,
        max_length=MAX_LEN,
        padding="max_length",
    )
    feed = {
        "input_ids": enc["input_ids"].astype(np.int64),
        "attention_mask": enc["attention_mask"].astype(np.int64),
    }
    logits_lid, logits_switch = session.run(None, feed)
    lid_id = int(logits_lid[0].argmax())
    switch_id = int(logits_switch[0].argmax())
    x = logits_lid[0].astype(np.float64)
    x = x - x.max()
    p = np.exp(x); p = p / p.sum()
    return {
        "text": text,
        "lid": id2lid[lid_id],
        "language_switch": id2switch[switch_id],
        "lid_confidence": float(p.max()),
    }

print(predict("Haan, ye sahi hai. Proceed karo."))

Output tensor names

Output	Shape	Head
`logits_lid`	`[batch, 10]`	Sentence LID
`logits_switch`	`[batch, 11]`	Language switch

Inputs: input_ids, attention_mask (int64, pad/truncate to 128).

No intent_logits output — use blue-machines/Intent-classifier-v1 for intent.

Downloads last month: -